Skip to content

Security: OneHillAI/ASDD

Security

SECURITY.md

Security Policy

ASDD's reference implementation runs AI agents against untrusted pull-request and issue content in CI. That is exactly the surface the 2026 attacks target, so we take reports seriously and the review workflow is the highest-scrutiny file in the repo.

Reporting a vulnerability

Please do not open a public issue for security vulnerabilities. Report it privately:

Please include a description and impact, steps to reproduce (a proof-of-concept if you have one), and the affected commit. We will acknowledge the report, keep you updated, and credit you when a fix ships unless you prefer to stay anonymous. Please allow reasonable time to fix before public disclosure.

In scope (especially welcome)

The review pipeline's threat model is documented in standards/security.md. Reports of any path that breaks these defenses are high priority:

  • A way for untrusted PR/issue content to reach an action-driving prompt (prompt-injection to action).
  • Any path where model output is executed as a command, or selects an arbitrary command.
  • The read-only analysis job gaining write scope or secret access, or the write-scoped publish job ingesting untrusted input.
  • Bypassing the policy decision point (scripts/policy-check.sh), getting an agent to merge, push, or take a non-allow-listed action.
  • A way around the protected-path human gate (CODEOWNERS + branch protection).
  • A supply-chain weakness in the workflows (unpinned action, a poisoned dependency path).

The script/workflow names above (policy-check.sh, the review workflows) refer to the reference implementation in this repository. This document defines the required properties; the implementation realizes them.

Adopter responsibility

If you adopt the reference implementation, review the two review workflows and the PDP before making your repo public, and run its conformance-check.sh. The defaults are safe, but you own your repository's configuration: branch protection, the CODEOWNERS owners, your protected_paths, and any runtime credential you add. A misconfigured adopter repo is the adopter's responsibility; a flaw in the shipped defaults is ours, report it.

There aren't any published security advisories